For autonomous robots to transition from structured factories to dynamic human environments, they must be fundamentally reliable. This thesis argues that reliability is not a monolithic property but must be managed along two distinct conceptual axes: Physical Risk and Task Risk. Physical Risk concerns the external danger of physical harm, such as collisions, arising from environmental uncertainty. Task Risk is the internal danger of mission failure, stemming from flaws in the robot's own generated plan. This dissertation develops principled, continuous-time frameworks to address challenges along both axes. Part I confronts the problem of estimating path-wise collision probability. Traditional discretization-based methods suffer from a fundamental dilemma: sparse checks lead to underestimation (“corner cutting”), while dense checks cause gross overestimation due to the mishandling of correlated events. This work addresses the issue by modeling risk as a continuous property of the path. We derive a path-wise functional, termed “Risk Density,” based on stochastic-process analysis, which provides an analytical, parameter-free measure of collision probability. This functional is embedded within RD-RRT*, a risk-aware motion planner that experimental validation shows can consistently find paths with demonstrably lower collision probability than standard approaches. Part II addresses a critical component of Task Risk: ensuring the correctness of behaviors learned by generative models. A robot that executes a flawed plan safely is not reliable. This challenge is magnified by modern policies that represent actions as discrete waypoint sequences, a choice that discards the essential dynamic information (e.g., velocity) required for rigorous analysis of stability or convergence. To address this limitation, we introduce a fundamental shift in representation, modeling actions as continuous-time functions within a Sobolev space (H¹). This guarantees “dynamic consistency,” ensuring learned policies are faithful to the complete dynamic state of demonstrated behaviors. Building on this foundation, we develop a differentiable certification pipeline. By repurposing a Control Lyapunov Function as an inference-time optimization objective, the framework can formally certify that a generated plan will make provable progress towards its goal, correcting statistically plausible but dynamically incorrect behaviors before execution. Scope and validation. Part I considers geometric motion planning problems in which collision risk arises from static environmental uncertainty modeled as Gaussian translational perturbations of workspace geometry. The proposed risk metric is designed for evaluating the path-wise probability of collision along a continuous trajectory, rather than propagating full state uncertainty through robot dynamics, and is used to produce open-loop plans with reduced collision probability. Part II considers manipulation behaviors represented as continuous-time trajectories, with task success specified as reaching a goal set in the world state. All results in this thesis are validated in simulation: Part I is evaluated by comparison to Monte Carlo estimates of collision probability and by planner benchmarks, and Part II is demonstrated on the PushT environment in a PyMunk-based simulator. In summary, this dissertation provides two principled, continuous-time methodologies for risk management in robotics. The Risk Density functional offers a robust solution to physical risk assessment in navigation, while the differentiable Lyapunov oracle establishes a framework for the certifiable execution of learned generative policies. Together, these methods support the development of autonomous systems designed to be both capable and quantitatively reliable.

RISK AWARE AND RELIABLE ROBOT AUTONOMY

PAIOLA, LORENZO
2026

Abstract

For autonomous robots to transition from structured factories to dynamic human environments, they must be fundamentally reliable. This thesis argues that reliability is not a monolithic property but must be managed along two distinct conceptual axes: Physical Risk and Task Risk. Physical Risk concerns the external danger of physical harm, such as collisions, arising from environmental uncertainty. Task Risk is the internal danger of mission failure, stemming from flaws in the robot's own generated plan. This dissertation develops principled, continuous-time frameworks to address challenges along both axes. Part I confronts the problem of estimating path-wise collision probability. Traditional discretization-based methods suffer from a fundamental dilemma: sparse checks lead to underestimation (“corner cutting”), while dense checks cause gross overestimation due to the mishandling of correlated events. This work addresses the issue by modeling risk as a continuous property of the path. We derive a path-wise functional, termed “Risk Density,” based on stochastic-process analysis, which provides an analytical, parameter-free measure of collision probability. This functional is embedded within RD-RRT*, a risk-aware motion planner that experimental validation shows can consistently find paths with demonstrably lower collision probability than standard approaches. Part II addresses a critical component of Task Risk: ensuring the correctness of behaviors learned by generative models. A robot that executes a flawed plan safely is not reliable. This challenge is magnified by modern policies that represent actions as discrete waypoint sequences, a choice that discards the essential dynamic information (e.g., velocity) required for rigorous analysis of stability or convergence. To address this limitation, we introduce a fundamental shift in representation, modeling actions as continuous-time functions within a Sobolev space (H¹). This guarantees “dynamic consistency,” ensuring learned policies are faithful to the complete dynamic state of demonstrated behaviors. Building on this foundation, we develop a differentiable certification pipeline. By repurposing a Control Lyapunov Function as an inference-time optimization objective, the framework can formally certify that a generated plan will make provable progress towards its goal, correcting statistically plausible but dynamically incorrect behaviors before execution. Scope and validation. Part I considers geometric motion planning problems in which collision risk arises from static environmental uncertainty modeled as Gaussian translational perturbations of workspace geometry. The proposed risk metric is designed for evaluating the path-wise probability of collision along a continuous trajectory, rather than propagating full state uncertainty through robot dynamics, and is used to produce open-loop plans with reduced collision probability. Part II considers manipulation behaviors represented as continuous-time trajectories, with task success specified as reaching a goal set in the world state. All results in this thesis are validated in simulation: Part I is evaluated by comparison to Monte Carlo estimates of collision probability and by planner benchmarks, and Part II is demonstrated on the PushT environment in a PyMunk-based simulator. In summary, this dissertation provides two principled, continuous-time methodologies for risk management in robotics. The Risk Density functional offers a robust solution to physical risk assessment in navigation, while the differentiable Lyapunov oracle establishes a framework for the certifiable execution of learned generative policies. Together, these methods support the development of autonomous systems designed to be both capable and quantitatively reliable.
6-set-2026
Inglese
generative modeling
planning
probability
risk
robotics
Bicchi, Antonio
Grioli, Giorgio
Catalano, Manuel Giuseppe
File in questo prodotto:
File Dimensione Formato  
mainthesis_pdfa.pdf

embargo fino al 08/09/2029

Licenza: Creative Commons
Dimensione 15.65 MB
Formato Adobe PDF
15.65 MB Adobe PDF

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/385759
Il codice NBN di questa tesi è URN:NBN:IT:UNIPI-385759